Ocfs2: Optimize punching-hole code.
[linux-2.6.git] / fs / ocfs2 / file.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * file.c
5  *
6  * File open, close, extend, truncate
7  *
8  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
9  *
10  * This program is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU General Public
12  * License as published by the Free Software Foundation; either
13  * version 2 of the License, or (at your option) any later version.
14  *
15  * This program is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * General Public License for more details.
19  *
20  * You should have received a copy of the GNU General Public
21  * License along with this program; if not, write to the
22  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23  * Boston, MA 021110-1307, USA.
24  */
25
26 #include <linux/capability.h>
27 #include <linux/fs.h>
28 #include <linux/types.h>
29 #include <linux/slab.h>
30 #include <linux/highmem.h>
31 #include <linux/pagemap.h>
32 #include <linux/uio.h>
33 #include <linux/sched.h>
34 #include <linux/splice.h>
35 #include <linux/mount.h>
36 #include <linux/writeback.h>
37 #include <linux/falloc.h>
38 #include <linux/quotaops.h>
39
40 #define MLOG_MASK_PREFIX ML_INODE
41 #include <cluster/masklog.h>
42
43 #include "ocfs2.h"
44
45 #include "alloc.h"
46 #include "aops.h"
47 #include "dir.h"
48 #include "dlmglue.h"
49 #include "extent_map.h"
50 #include "file.h"
51 #include "sysfile.h"
52 #include "inode.h"
53 #include "ioctl.h"
54 #include "journal.h"
55 #include "locks.h"
56 #include "mmap.h"
57 #include "suballoc.h"
58 #include "super.h"
59 #include "xattr.h"
60 #include "acl.h"
61 #include "quota.h"
62 #include "refcounttree.h"
63
64 #include "buffer_head_io.h"
65
66 static int ocfs2_sync_inode(struct inode *inode)
67 {
68         filemap_fdatawrite(inode->i_mapping);
69         return sync_mapping_buffers(inode->i_mapping);
70 }
71
72 static int ocfs2_init_file_private(struct inode *inode, struct file *file)
73 {
74         struct ocfs2_file_private *fp;
75
76         fp = kzalloc(sizeof(struct ocfs2_file_private), GFP_KERNEL);
77         if (!fp)
78                 return -ENOMEM;
79
80         fp->fp_file = file;
81         mutex_init(&fp->fp_mutex);
82         ocfs2_file_lock_res_init(&fp->fp_flock, fp);
83         file->private_data = fp;
84
85         return 0;
86 }
87
88 static void ocfs2_free_file_private(struct inode *inode, struct file *file)
89 {
90         struct ocfs2_file_private *fp = file->private_data;
91         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
92
93         if (fp) {
94                 ocfs2_simple_drop_lockres(osb, &fp->fp_flock);
95                 ocfs2_lock_res_free(&fp->fp_flock);
96                 kfree(fp);
97                 file->private_data = NULL;
98         }
99 }
100
101 static int ocfs2_file_open(struct inode *inode, struct file *file)
102 {
103         int status;
104         int mode = file->f_flags;
105         struct ocfs2_inode_info *oi = OCFS2_I(inode);
106
107         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
108                    file->f_path.dentry->d_name.len, file->f_path.dentry->d_name.name);
109
110         if (file->f_mode & FMODE_WRITE)
111                 dquot_initialize(inode);
112
113         spin_lock(&oi->ip_lock);
114
115         /* Check that the inode hasn't been wiped from disk by another
116          * node. If it hasn't then we're safe as long as we hold the
117          * spin lock until our increment of open count. */
118         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
119                 spin_unlock(&oi->ip_lock);
120
121                 status = -ENOENT;
122                 goto leave;
123         }
124
125         if (mode & O_DIRECT)
126                 oi->ip_flags |= OCFS2_INODE_OPEN_DIRECT;
127
128         oi->ip_open_count++;
129         spin_unlock(&oi->ip_lock);
130
131         status = ocfs2_init_file_private(inode, file);
132         if (status) {
133                 /*
134                  * We want to set open count back if we're failing the
135                  * open.
136                  */
137                 spin_lock(&oi->ip_lock);
138                 oi->ip_open_count--;
139                 spin_unlock(&oi->ip_lock);
140         }
141
142 leave:
143         mlog_exit(status);
144         return status;
145 }
146
147 static int ocfs2_file_release(struct inode *inode, struct file *file)
148 {
149         struct ocfs2_inode_info *oi = OCFS2_I(inode);
150
151         mlog_entry("(0x%p, 0x%p, '%.*s')\n", inode, file,
152                        file->f_path.dentry->d_name.len,
153                        file->f_path.dentry->d_name.name);
154
155         spin_lock(&oi->ip_lock);
156         if (!--oi->ip_open_count)
157                 oi->ip_flags &= ~OCFS2_INODE_OPEN_DIRECT;
158         spin_unlock(&oi->ip_lock);
159
160         ocfs2_free_file_private(inode, file);
161
162         mlog_exit(0);
163
164         return 0;
165 }
166
167 static int ocfs2_dir_open(struct inode *inode, struct file *file)
168 {
169         return ocfs2_init_file_private(inode, file);
170 }
171
172 static int ocfs2_dir_release(struct inode *inode, struct file *file)
173 {
174         ocfs2_free_file_private(inode, file);
175         return 0;
176 }
177
178 static int ocfs2_sync_file(struct file *file,
179                            struct dentry *dentry,
180                            int datasync)
181 {
182         int err = 0;
183         journal_t *journal;
184         struct inode *inode = dentry->d_inode;
185         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
186
187         mlog_entry("(0x%p, 0x%p, %d, '%.*s')\n", file, dentry, datasync,
188                    dentry->d_name.len, dentry->d_name.name);
189
190         err = ocfs2_sync_inode(dentry->d_inode);
191         if (err)
192                 goto bail;
193
194         if (datasync && !(inode->i_state & I_DIRTY_DATASYNC))
195                 goto bail;
196
197         journal = osb->journal->j_journal;
198         err = jbd2_journal_force_commit(journal);
199
200 bail:
201         mlog_exit(err);
202
203         return (err < 0) ? -EIO : 0;
204 }
205
206 int ocfs2_should_update_atime(struct inode *inode,
207                               struct vfsmount *vfsmnt)
208 {
209         struct timespec now;
210         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
211
212         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
213                 return 0;
214
215         if ((inode->i_flags & S_NOATIME) ||
216             ((inode->i_sb->s_flags & MS_NODIRATIME) && S_ISDIR(inode->i_mode)))
217                 return 0;
218
219         /*
220          * We can be called with no vfsmnt structure - NFSD will
221          * sometimes do this.
222          *
223          * Note that our action here is different than touch_atime() -
224          * if we can't tell whether this is a noatime mount, then we
225          * don't know whether to trust the value of s_atime_quantum.
226          */
227         if (vfsmnt == NULL)
228                 return 0;
229
230         if ((vfsmnt->mnt_flags & MNT_NOATIME) ||
231             ((vfsmnt->mnt_flags & MNT_NODIRATIME) && S_ISDIR(inode->i_mode)))
232                 return 0;
233
234         if (vfsmnt->mnt_flags & MNT_RELATIME) {
235                 if ((timespec_compare(&inode->i_atime, &inode->i_mtime) <= 0) ||
236                     (timespec_compare(&inode->i_atime, &inode->i_ctime) <= 0))
237                         return 1;
238
239                 return 0;
240         }
241
242         now = CURRENT_TIME;
243         if ((now.tv_sec - inode->i_atime.tv_sec <= osb->s_atime_quantum))
244                 return 0;
245         else
246                 return 1;
247 }
248
249 int ocfs2_update_inode_atime(struct inode *inode,
250                              struct buffer_head *bh)
251 {
252         int ret;
253         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
254         handle_t *handle;
255         struct ocfs2_dinode *di = (struct ocfs2_dinode *) bh->b_data;
256
257         mlog_entry_void();
258
259         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
260         if (IS_ERR(handle)) {
261                 ret = PTR_ERR(handle);
262                 mlog_errno(ret);
263                 goto out;
264         }
265
266         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
267                                       OCFS2_JOURNAL_ACCESS_WRITE);
268         if (ret) {
269                 mlog_errno(ret);
270                 goto out_commit;
271         }
272
273         /*
274          * Don't use ocfs2_mark_inode_dirty() here as we don't always
275          * have i_mutex to guard against concurrent changes to other
276          * inode fields.
277          */
278         inode->i_atime = CURRENT_TIME;
279         di->i_atime = cpu_to_le64(inode->i_atime.tv_sec);
280         di->i_atime_nsec = cpu_to_le32(inode->i_atime.tv_nsec);
281         ocfs2_journal_dirty(handle, bh);
282
283 out_commit:
284         ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
285 out:
286         mlog_exit(ret);
287         return ret;
288 }
289
290 static int ocfs2_set_inode_size(handle_t *handle,
291                                 struct inode *inode,
292                                 struct buffer_head *fe_bh,
293                                 u64 new_i_size)
294 {
295         int status;
296
297         mlog_entry_void();
298         i_size_write(inode, new_i_size);
299         inode->i_blocks = ocfs2_inode_sector_count(inode);
300         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
301
302         status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
303         if (status < 0) {
304                 mlog_errno(status);
305                 goto bail;
306         }
307
308 bail:
309         mlog_exit(status);
310         return status;
311 }
312
313 int ocfs2_simple_size_update(struct inode *inode,
314                              struct buffer_head *di_bh,
315                              u64 new_i_size)
316 {
317         int ret;
318         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
319         handle_t *handle = NULL;
320
321         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
322         if (IS_ERR(handle)) {
323                 ret = PTR_ERR(handle);
324                 mlog_errno(ret);
325                 goto out;
326         }
327
328         ret = ocfs2_set_inode_size(handle, inode, di_bh,
329                                    new_i_size);
330         if (ret < 0)
331                 mlog_errno(ret);
332
333         ocfs2_commit_trans(osb, handle);
334 out:
335         return ret;
336 }
337
338 static int ocfs2_cow_file_pos(struct inode *inode,
339                               struct buffer_head *fe_bh,
340                               u64 offset)
341 {
342         int status;
343         u32 phys, cpos = offset >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
344         unsigned int num_clusters = 0;
345         unsigned int ext_flags = 0;
346
347         /*
348          * If the new offset is aligned to the range of the cluster, there is
349          * no space for ocfs2_zero_range_for_truncate to fill, so no need to
350          * CoW either.
351          */
352         if ((offset & (OCFS2_SB(inode->i_sb)->s_clustersize - 1)) == 0)
353                 return 0;
354
355         status = ocfs2_get_clusters(inode, cpos, &phys,
356                                     &num_clusters, &ext_flags);
357         if (status) {
358                 mlog_errno(status);
359                 goto out;
360         }
361
362         if (!(ext_flags & OCFS2_EXT_REFCOUNTED))
363                 goto out;
364
365         return ocfs2_refcount_cow(inode, fe_bh, cpos, 1, cpos+1);
366
367 out:
368         return status;
369 }
370
371 static int ocfs2_orphan_for_truncate(struct ocfs2_super *osb,
372                                      struct inode *inode,
373                                      struct buffer_head *fe_bh,
374                                      u64 new_i_size)
375 {
376         int status;
377         handle_t *handle;
378         struct ocfs2_dinode *di;
379         u64 cluster_bytes;
380
381         mlog_entry_void();
382
383         /*
384          * We need to CoW the cluster contains the offset if it is reflinked
385          * since we will call ocfs2_zero_range_for_truncate later which will
386          * write "0" from offset to the end of the cluster.
387          */
388         status = ocfs2_cow_file_pos(inode, fe_bh, new_i_size);
389         if (status) {
390                 mlog_errno(status);
391                 return status;
392         }
393
394         /* TODO: This needs to actually orphan the inode in this
395          * transaction. */
396
397         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
398         if (IS_ERR(handle)) {
399                 status = PTR_ERR(handle);
400                 mlog_errno(status);
401                 goto out;
402         }
403
404         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), fe_bh,
405                                          OCFS2_JOURNAL_ACCESS_WRITE);
406         if (status < 0) {
407                 mlog_errno(status);
408                 goto out_commit;
409         }
410
411         /*
412          * Do this before setting i_size.
413          */
414         cluster_bytes = ocfs2_align_bytes_to_clusters(inode->i_sb, new_i_size);
415         status = ocfs2_zero_range_for_truncate(inode, handle, new_i_size,
416                                                cluster_bytes);
417         if (status) {
418                 mlog_errno(status);
419                 goto out_commit;
420         }
421
422         i_size_write(inode, new_i_size);
423         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
424
425         di = (struct ocfs2_dinode *) fe_bh->b_data;
426         di->i_size = cpu_to_le64(new_i_size);
427         di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec);
428         di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec);
429
430         ocfs2_journal_dirty(handle, fe_bh);
431
432 out_commit:
433         ocfs2_commit_trans(osb, handle);
434 out:
435
436         mlog_exit(status);
437         return status;
438 }
439
440 static int ocfs2_truncate_file(struct inode *inode,
441                                struct buffer_head *di_bh,
442                                u64 new_i_size)
443 {
444         int status = 0;
445         struct ocfs2_dinode *fe = NULL;
446         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
447
448         mlog_entry("(inode = %llu, new_i_size = %llu\n",
449                    (unsigned long long)OCFS2_I(inode)->ip_blkno,
450                    (unsigned long long)new_i_size);
451
452         /* We trust di_bh because it comes from ocfs2_inode_lock(), which
453          * already validated it */
454         fe = (struct ocfs2_dinode *) di_bh->b_data;
455
456         mlog_bug_on_msg(le64_to_cpu(fe->i_size) != i_size_read(inode),
457                         "Inode %llu, inode i_size = %lld != di "
458                         "i_size = %llu, i_flags = 0x%x\n",
459                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
460                         i_size_read(inode),
461                         (unsigned long long)le64_to_cpu(fe->i_size),
462                         le32_to_cpu(fe->i_flags));
463
464         if (new_i_size > le64_to_cpu(fe->i_size)) {
465                 mlog(0, "asked to truncate file with size (%llu) to size (%llu)!\n",
466                      (unsigned long long)le64_to_cpu(fe->i_size),
467                      (unsigned long long)new_i_size);
468                 status = -EINVAL;
469                 mlog_errno(status);
470                 goto bail;
471         }
472
473         mlog(0, "inode %llu, i_size = %llu, new_i_size = %llu\n",
474              (unsigned long long)le64_to_cpu(fe->i_blkno),
475              (unsigned long long)le64_to_cpu(fe->i_size),
476              (unsigned long long)new_i_size);
477
478         /* lets handle the simple truncate cases before doing any more
479          * cluster locking. */
480         if (new_i_size == le64_to_cpu(fe->i_size))
481                 goto bail;
482
483         down_write(&OCFS2_I(inode)->ip_alloc_sem);
484
485         ocfs2_resv_discard(&osb->osb_la_resmap,
486                            &OCFS2_I(inode)->ip_la_data_resv);
487
488         /*
489          * The inode lock forced other nodes to sync and drop their
490          * pages, which (correctly) happens even if we have a truncate
491          * without allocation change - ocfs2 cluster sizes can be much
492          * greater than page size, so we have to truncate them
493          * anyway.
494          */
495         unmap_mapping_range(inode->i_mapping, new_i_size + PAGE_SIZE - 1, 0, 1);
496         truncate_inode_pages(inode->i_mapping, new_i_size);
497
498         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
499                 status = ocfs2_truncate_inline(inode, di_bh, new_i_size,
500                                                i_size_read(inode), 1);
501                 if (status)
502                         mlog_errno(status);
503
504                 goto bail_unlock_sem;
505         }
506
507         /* alright, we're going to need to do a full blown alloc size
508          * change. Orphan the inode so that recovery can complete the
509          * truncate if necessary. This does the task of marking
510          * i_size. */
511         status = ocfs2_orphan_for_truncate(osb, inode, di_bh, new_i_size);
512         if (status < 0) {
513                 mlog_errno(status);
514                 goto bail_unlock_sem;
515         }
516
517         status = ocfs2_commit_truncate(osb, inode, di_bh);
518         if (status < 0) {
519                 mlog_errno(status);
520                 goto bail_unlock_sem;
521         }
522
523         /* TODO: orphan dir cleanup here. */
524 bail_unlock_sem:
525         up_write(&OCFS2_I(inode)->ip_alloc_sem);
526
527 bail:
528         if (!status && OCFS2_I(inode)->ip_clusters == 0)
529                 status = ocfs2_try_remove_refcount_tree(inode, di_bh);
530
531         mlog_exit(status);
532         return status;
533 }
534
535 /*
536  * extend file allocation only here.
537  * we'll update all the disk stuff, and oip->alloc_size
538  *
539  * expect stuff to be locked, a transaction started and enough data /
540  * metadata reservations in the contexts.
541  *
542  * Will return -EAGAIN, and a reason if a restart is needed.
543  * If passed in, *reason will always be set, even in error.
544  */
545 int ocfs2_add_inode_data(struct ocfs2_super *osb,
546                          struct inode *inode,
547                          u32 *logical_offset,
548                          u32 clusters_to_add,
549                          int mark_unwritten,
550                          struct buffer_head *fe_bh,
551                          handle_t *handle,
552                          struct ocfs2_alloc_context *data_ac,
553                          struct ocfs2_alloc_context *meta_ac,
554                          enum ocfs2_alloc_restarted *reason_ret)
555 {
556         int ret;
557         struct ocfs2_extent_tree et;
558
559         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), fe_bh);
560         ret = ocfs2_add_clusters_in_btree(handle, &et, logical_offset,
561                                           clusters_to_add, mark_unwritten,
562                                           data_ac, meta_ac, reason_ret);
563
564         return ret;
565 }
566
567 static int __ocfs2_extend_allocation(struct inode *inode, u32 logical_start,
568                                      u32 clusters_to_add, int mark_unwritten)
569 {
570         int status = 0;
571         int restart_func = 0;
572         int credits;
573         u32 prev_clusters;
574         struct buffer_head *bh = NULL;
575         struct ocfs2_dinode *fe = NULL;
576         handle_t *handle = NULL;
577         struct ocfs2_alloc_context *data_ac = NULL;
578         struct ocfs2_alloc_context *meta_ac = NULL;
579         enum ocfs2_alloc_restarted why;
580         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
581         struct ocfs2_extent_tree et;
582         int did_quota = 0;
583
584         mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
585
586         /*
587          * This function only exists for file systems which don't
588          * support holes.
589          */
590         BUG_ON(mark_unwritten && !ocfs2_sparse_alloc(osb));
591
592         status = ocfs2_read_inode_block(inode, &bh);
593         if (status < 0) {
594                 mlog_errno(status);
595                 goto leave;
596         }
597         fe = (struct ocfs2_dinode *) bh->b_data;
598
599 restart_all:
600         BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
601
602         mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u, "
603              "clusters_to_add = %u\n",
604              (unsigned long long)OCFS2_I(inode)->ip_blkno,
605              (long long)i_size_read(inode), le32_to_cpu(fe->i_clusters),
606              clusters_to_add);
607         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), bh);
608         status = ocfs2_lock_allocators(inode, &et, clusters_to_add, 0,
609                                        &data_ac, &meta_ac);
610         if (status) {
611                 mlog_errno(status);
612                 goto leave;
613         }
614
615         credits = ocfs2_calc_extend_credits(osb->sb, &fe->id2.i_list,
616                                             clusters_to_add);
617         handle = ocfs2_start_trans(osb, credits);
618         if (IS_ERR(handle)) {
619                 status = PTR_ERR(handle);
620                 handle = NULL;
621                 mlog_errno(status);
622                 goto leave;
623         }
624
625 restarted_transaction:
626         status = dquot_alloc_space_nodirty(inode,
627                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
628         if (status)
629                 goto leave;
630         did_quota = 1;
631
632         /* reserve a write to the file entry early on - that we if we
633          * run out of credits in the allocation path, we can still
634          * update i_size. */
635         status = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
636                                          OCFS2_JOURNAL_ACCESS_WRITE);
637         if (status < 0) {
638                 mlog_errno(status);
639                 goto leave;
640         }
641
642         prev_clusters = OCFS2_I(inode)->ip_clusters;
643
644         status = ocfs2_add_inode_data(osb,
645                                       inode,
646                                       &logical_start,
647                                       clusters_to_add,
648                                       mark_unwritten,
649                                       bh,
650                                       handle,
651                                       data_ac,
652                                       meta_ac,
653                                       &why);
654         if ((status < 0) && (status != -EAGAIN)) {
655                 if (status != -ENOSPC)
656                         mlog_errno(status);
657                 goto leave;
658         }
659
660         ocfs2_journal_dirty(handle, bh);
661
662         spin_lock(&OCFS2_I(inode)->ip_lock);
663         clusters_to_add -= (OCFS2_I(inode)->ip_clusters - prev_clusters);
664         spin_unlock(&OCFS2_I(inode)->ip_lock);
665         /* Release unused quota reservation */
666         dquot_free_space(inode,
667                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
668         did_quota = 0;
669
670         if (why != RESTART_NONE && clusters_to_add) {
671                 if (why == RESTART_META) {
672                         mlog(0, "restarting function.\n");
673                         restart_func = 1;
674                 } else {
675                         BUG_ON(why != RESTART_TRANS);
676
677                         mlog(0, "restarting transaction.\n");
678                         /* TODO: This can be more intelligent. */
679                         credits = ocfs2_calc_extend_credits(osb->sb,
680                                                             &fe->id2.i_list,
681                                                             clusters_to_add);
682                         status = ocfs2_extend_trans(handle, credits);
683                         if (status < 0) {
684                                 /* handle still has to be committed at
685                                  * this point. */
686                                 status = -ENOMEM;
687                                 mlog_errno(status);
688                                 goto leave;
689                         }
690                         goto restarted_transaction;
691                 }
692         }
693
694         mlog(0, "fe: i_clusters = %u, i_size=%llu\n",
695              le32_to_cpu(fe->i_clusters),
696              (unsigned long long)le64_to_cpu(fe->i_size));
697         mlog(0, "inode: ip_clusters=%u, i_size=%lld\n",
698              OCFS2_I(inode)->ip_clusters, (long long)i_size_read(inode));
699
700 leave:
701         if (status < 0 && did_quota)
702                 dquot_free_space(inode,
703                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_add));
704         if (handle) {
705                 ocfs2_commit_trans(osb, handle);
706                 handle = NULL;
707         }
708         if (data_ac) {
709                 ocfs2_free_alloc_context(data_ac);
710                 data_ac = NULL;
711         }
712         if (meta_ac) {
713                 ocfs2_free_alloc_context(meta_ac);
714                 meta_ac = NULL;
715         }
716         if ((!status) && restart_func) {
717                 restart_func = 0;
718                 goto restart_all;
719         }
720         brelse(bh);
721         bh = NULL;
722
723         mlog_exit(status);
724         return status;
725 }
726
727 /* Some parts of this taken from generic_cont_expand, which turned out
728  * to be too fragile to do exactly what we need without us having to
729  * worry about recursive locking in ->write_begin() and ->write_end(). */
730 static int ocfs2_write_zero_page(struct inode *inode,
731                                  u64 size)
732 {
733         struct address_space *mapping = inode->i_mapping;
734         struct page *page;
735         unsigned long index;
736         unsigned int offset;
737         handle_t *handle = NULL;
738         int ret;
739
740         offset = (size & (PAGE_CACHE_SIZE-1)); /* Within page */
741         /* ugh.  in prepare/commit_write, if from==to==start of block, we
742         ** skip the prepare.  make sure we never send an offset for the start
743         ** of a block
744         */
745         if ((offset & (inode->i_sb->s_blocksize - 1)) == 0) {
746                 offset++;
747         }
748         index = size >> PAGE_CACHE_SHIFT;
749
750         page = grab_cache_page(mapping, index);
751         if (!page) {
752                 ret = -ENOMEM;
753                 mlog_errno(ret);
754                 goto out;
755         }
756
757         ret = ocfs2_prepare_write_nolock(inode, page, offset, offset);
758         if (ret < 0) {
759                 mlog_errno(ret);
760                 goto out_unlock;
761         }
762
763         if (ocfs2_should_order_data(inode)) {
764                 handle = ocfs2_start_walk_page_trans(inode, page, offset,
765                                                      offset);
766                 if (IS_ERR(handle)) {
767                         ret = PTR_ERR(handle);
768                         handle = NULL;
769                         goto out_unlock;
770                 }
771         }
772
773         /* must not update i_size! */
774         ret = block_commit_write(page, offset, offset);
775         if (ret < 0)
776                 mlog_errno(ret);
777         else
778                 ret = 0;
779
780         if (handle)
781                 ocfs2_commit_trans(OCFS2_SB(inode->i_sb), handle);
782 out_unlock:
783         unlock_page(page);
784         page_cache_release(page);
785 out:
786         return ret;
787 }
788
789 static int ocfs2_zero_extend(struct inode *inode,
790                              u64 zero_to_size)
791 {
792         int ret = 0;
793         u64 start_off;
794         struct super_block *sb = inode->i_sb;
795
796         start_off = ocfs2_align_bytes_to_blocks(sb, i_size_read(inode));
797         while (start_off < zero_to_size) {
798                 ret = ocfs2_write_zero_page(inode, start_off);
799                 if (ret < 0) {
800                         mlog_errno(ret);
801                         goto out;
802                 }
803
804                 start_off += sb->s_blocksize;
805
806                 /*
807                  * Very large extends have the potential to lock up
808                  * the cpu for extended periods of time.
809                  */
810                 cond_resched();
811         }
812
813 out:
814         return ret;
815 }
816
817 int ocfs2_extend_no_holes(struct inode *inode, u64 new_i_size, u64 zero_to)
818 {
819         int ret;
820         u32 clusters_to_add;
821         struct ocfs2_inode_info *oi = OCFS2_I(inode);
822
823         clusters_to_add = ocfs2_clusters_for_bytes(inode->i_sb, new_i_size);
824         if (clusters_to_add < oi->ip_clusters)
825                 clusters_to_add = 0;
826         else
827                 clusters_to_add -= oi->ip_clusters;
828
829         if (clusters_to_add) {
830                 ret = __ocfs2_extend_allocation(inode, oi->ip_clusters,
831                                                 clusters_to_add, 0);
832                 if (ret) {
833                         mlog_errno(ret);
834                         goto out;
835                 }
836         }
837
838         /*
839          * Call this even if we don't add any clusters to the tree. We
840          * still need to zero the area between the old i_size and the
841          * new i_size.
842          */
843         ret = ocfs2_zero_extend(inode, zero_to);
844         if (ret < 0)
845                 mlog_errno(ret);
846
847 out:
848         return ret;
849 }
850
851 static int ocfs2_extend_file(struct inode *inode,
852                              struct buffer_head *di_bh,
853                              u64 new_i_size)
854 {
855         int ret = 0;
856         struct ocfs2_inode_info *oi = OCFS2_I(inode);
857
858         BUG_ON(!di_bh);
859
860         /* setattr sometimes calls us like this. */
861         if (new_i_size == 0)
862                 goto out;
863
864         if (i_size_read(inode) == new_i_size)
865                 goto out;
866         BUG_ON(new_i_size < i_size_read(inode));
867
868         /*
869          * Fall through for converting inline data, even if the fs
870          * supports sparse files.
871          *
872          * The check for inline data here is legal - nobody can add
873          * the feature since we have i_mutex. We must check it again
874          * after acquiring ip_alloc_sem though, as paths like mmap
875          * might have raced us to converting the inode to extents.
876          */
877         if (!(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
878             && ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
879                 goto out_update_size;
880
881         /*
882          * The alloc sem blocks people in read/write from reading our
883          * allocation until we're done changing it. We depend on
884          * i_mutex to block other extend/truncate calls while we're
885          * here.
886          */
887         down_write(&oi->ip_alloc_sem);
888
889         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
890                 /*
891                  * We can optimize small extends by keeping the inodes
892                  * inline data.
893                  */
894                 if (ocfs2_size_fits_inline_data(di_bh, new_i_size)) {
895                         up_write(&oi->ip_alloc_sem);
896                         goto out_update_size;
897                 }
898
899                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
900                 if (ret) {
901                         up_write(&oi->ip_alloc_sem);
902
903                         mlog_errno(ret);
904                         goto out;
905                 }
906         }
907
908         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
909                 ret = ocfs2_extend_no_holes(inode, new_i_size, new_i_size);
910
911         up_write(&oi->ip_alloc_sem);
912
913         if (ret < 0) {
914                 mlog_errno(ret);
915                 goto out;
916         }
917
918 out_update_size:
919         ret = ocfs2_simple_size_update(inode, di_bh, new_i_size);
920         if (ret < 0)
921                 mlog_errno(ret);
922
923 out:
924         return ret;
925 }
926
927 int ocfs2_setattr(struct dentry *dentry, struct iattr *attr)
928 {
929         int status = 0, size_change;
930         struct inode *inode = dentry->d_inode;
931         struct super_block *sb = inode->i_sb;
932         struct ocfs2_super *osb = OCFS2_SB(sb);
933         struct buffer_head *bh = NULL;
934         handle_t *handle = NULL;
935         int qtype;
936         struct dquot *transfer_from[MAXQUOTAS] = { };
937         struct dquot *transfer_to[MAXQUOTAS] = { };
938
939         mlog_entry("(0x%p, '%.*s')\n", dentry,
940                    dentry->d_name.len, dentry->d_name.name);
941
942         /* ensuring we don't even attempt to truncate a symlink */
943         if (S_ISLNK(inode->i_mode))
944                 attr->ia_valid &= ~ATTR_SIZE;
945
946         if (attr->ia_valid & ATTR_MODE)
947                 mlog(0, "mode change: %d\n", attr->ia_mode);
948         if (attr->ia_valid & ATTR_UID)
949                 mlog(0, "uid change: %d\n", attr->ia_uid);
950         if (attr->ia_valid & ATTR_GID)
951                 mlog(0, "gid change: %d\n", attr->ia_gid);
952         if (attr->ia_valid & ATTR_SIZE)
953                 mlog(0, "size change...\n");
954         if (attr->ia_valid & (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME))
955                 mlog(0, "time change...\n");
956
957 #define OCFS2_VALID_ATTRS (ATTR_ATIME | ATTR_MTIME | ATTR_CTIME | ATTR_SIZE \
958                            | ATTR_GID | ATTR_UID | ATTR_MODE)
959         if (!(attr->ia_valid & OCFS2_VALID_ATTRS)) {
960                 mlog(0, "can't handle attrs: 0x%x\n", attr->ia_valid);
961                 return 0;
962         }
963
964         status = inode_change_ok(inode, attr);
965         if (status)
966                 return status;
967
968         size_change = S_ISREG(inode->i_mode) && attr->ia_valid & ATTR_SIZE;
969         if (size_change) {
970                 dquot_initialize(inode);
971
972                 status = ocfs2_rw_lock(inode, 1);
973                 if (status < 0) {
974                         mlog_errno(status);
975                         goto bail;
976                 }
977         }
978
979         status = ocfs2_inode_lock(inode, &bh, 1);
980         if (status < 0) {
981                 if (status != -ENOENT)
982                         mlog_errno(status);
983                 goto bail_unlock_rw;
984         }
985
986         if (size_change && attr->ia_size != i_size_read(inode)) {
987                 status = inode_newsize_ok(inode, attr->ia_size);
988                 if (status)
989                         goto bail_unlock;
990
991                 if (i_size_read(inode) > attr->ia_size) {
992                         if (ocfs2_should_order_data(inode)) {
993                                 status = ocfs2_begin_ordered_truncate(inode,
994                                                                       attr->ia_size);
995                                 if (status)
996                                         goto bail_unlock;
997                         }
998                         status = ocfs2_truncate_file(inode, bh, attr->ia_size);
999                 } else
1000                         status = ocfs2_extend_file(inode, bh, attr->ia_size);
1001                 if (status < 0) {
1002                         if (status != -ENOSPC)
1003                                 mlog_errno(status);
1004                         status = -ENOSPC;
1005                         goto bail_unlock;
1006                 }
1007         }
1008
1009         if ((attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid) ||
1010             (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid)) {
1011                 /*
1012                  * Gather pointers to quota structures so that allocation /
1013                  * freeing of quota structures happens here and not inside
1014                  * dquot_transfer() where we have problems with lock ordering
1015                  */
1016                 if (attr->ia_valid & ATTR_UID && attr->ia_uid != inode->i_uid
1017                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1018                     OCFS2_FEATURE_RO_COMPAT_USRQUOTA)) {
1019                         transfer_to[USRQUOTA] = dqget(sb, attr->ia_uid,
1020                                                       USRQUOTA);
1021                         transfer_from[USRQUOTA] = dqget(sb, inode->i_uid,
1022                                                         USRQUOTA);
1023                         if (!transfer_to[USRQUOTA] || !transfer_from[USRQUOTA]) {
1024                                 status = -ESRCH;
1025                                 goto bail_unlock;
1026                         }
1027                 }
1028                 if (attr->ia_valid & ATTR_GID && attr->ia_gid != inode->i_gid
1029                     && OCFS2_HAS_RO_COMPAT_FEATURE(sb,
1030                     OCFS2_FEATURE_RO_COMPAT_GRPQUOTA)) {
1031                         transfer_to[GRPQUOTA] = dqget(sb, attr->ia_gid,
1032                                                       GRPQUOTA);
1033                         transfer_from[GRPQUOTA] = dqget(sb, inode->i_gid,
1034                                                         GRPQUOTA);
1035                         if (!transfer_to[GRPQUOTA] || !transfer_from[GRPQUOTA]) {
1036                                 status = -ESRCH;
1037                                 goto bail_unlock;
1038                         }
1039                 }
1040                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS +
1041                                            2 * ocfs2_quota_trans_credits(sb));
1042                 if (IS_ERR(handle)) {
1043                         status = PTR_ERR(handle);
1044                         mlog_errno(status);
1045                         goto bail_unlock;
1046                 }
1047                 status = dquot_transfer(inode, attr);
1048                 if (status < 0)
1049                         goto bail_commit;
1050         } else {
1051                 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1052                 if (IS_ERR(handle)) {
1053                         status = PTR_ERR(handle);
1054                         mlog_errno(status);
1055                         goto bail_unlock;
1056                 }
1057         }
1058
1059         /*
1060          * This will intentionally not wind up calling vmtruncate(),
1061          * since all the work for a size change has been done above.
1062          * Otherwise, we could get into problems with truncate as
1063          * ip_alloc_sem is used there to protect against i_size
1064          * changes.
1065          */
1066         status = inode_setattr(inode, attr);
1067         if (status < 0) {
1068                 mlog_errno(status);
1069                 goto bail_commit;
1070         }
1071
1072         status = ocfs2_mark_inode_dirty(handle, inode, bh);
1073         if (status < 0)
1074                 mlog_errno(status);
1075
1076 bail_commit:
1077         ocfs2_commit_trans(osb, handle);
1078 bail_unlock:
1079         ocfs2_inode_unlock(inode, 1);
1080 bail_unlock_rw:
1081         if (size_change)
1082                 ocfs2_rw_unlock(inode, 1);
1083 bail:
1084         brelse(bh);
1085
1086         /* Release quota pointers in case we acquired them */
1087         for (qtype = 0; qtype < MAXQUOTAS; qtype++) {
1088                 dqput(transfer_to[qtype]);
1089                 dqput(transfer_from[qtype]);
1090         }
1091
1092         if (!status && attr->ia_valid & ATTR_MODE) {
1093                 status = ocfs2_acl_chmod(inode);
1094                 if (status < 0)
1095                         mlog_errno(status);
1096         }
1097
1098         mlog_exit(status);
1099         return status;
1100 }
1101
1102 int ocfs2_getattr(struct vfsmount *mnt,
1103                   struct dentry *dentry,
1104                   struct kstat *stat)
1105 {
1106         struct inode *inode = dentry->d_inode;
1107         struct super_block *sb = dentry->d_inode->i_sb;
1108         struct ocfs2_super *osb = sb->s_fs_info;
1109         int err;
1110
1111         mlog_entry_void();
1112
1113         err = ocfs2_inode_revalidate(dentry);
1114         if (err) {
1115                 if (err != -ENOENT)
1116                         mlog_errno(err);
1117                 goto bail;
1118         }
1119
1120         generic_fillattr(inode, stat);
1121
1122         /* We set the blksize from the cluster size for performance */
1123         stat->blksize = osb->s_clustersize;
1124
1125 bail:
1126         mlog_exit(err);
1127
1128         return err;
1129 }
1130
1131 int ocfs2_permission(struct inode *inode, int mask)
1132 {
1133         int ret;
1134
1135         mlog_entry_void();
1136
1137         ret = ocfs2_inode_lock(inode, NULL, 0);
1138         if (ret) {
1139                 if (ret != -ENOENT)
1140                         mlog_errno(ret);
1141                 goto out;
1142         }
1143
1144         ret = generic_permission(inode, mask, ocfs2_check_acl);
1145
1146         ocfs2_inode_unlock(inode, 0);
1147 out:
1148         mlog_exit(ret);
1149         return ret;
1150 }
1151
1152 static int __ocfs2_write_remove_suid(struct inode *inode,
1153                                      struct buffer_head *bh)
1154 {
1155         int ret;
1156         handle_t *handle;
1157         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1158         struct ocfs2_dinode *di;
1159
1160         mlog_entry("(Inode %llu, mode 0%o)\n",
1161                    (unsigned long long)OCFS2_I(inode)->ip_blkno, inode->i_mode);
1162
1163         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1164         if (IS_ERR(handle)) {
1165                 ret = PTR_ERR(handle);
1166                 mlog_errno(ret);
1167                 goto out;
1168         }
1169
1170         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), bh,
1171                                       OCFS2_JOURNAL_ACCESS_WRITE);
1172         if (ret < 0) {
1173                 mlog_errno(ret);
1174                 goto out_trans;
1175         }
1176
1177         inode->i_mode &= ~S_ISUID;
1178         if ((inode->i_mode & S_ISGID) && (inode->i_mode & S_IXGRP))
1179                 inode->i_mode &= ~S_ISGID;
1180
1181         di = (struct ocfs2_dinode *) bh->b_data;
1182         di->i_mode = cpu_to_le16(inode->i_mode);
1183
1184         ocfs2_journal_dirty(handle, bh);
1185
1186 out_trans:
1187         ocfs2_commit_trans(osb, handle);
1188 out:
1189         mlog_exit(ret);
1190         return ret;
1191 }
1192
1193 /*
1194  * Will look for holes and unwritten extents in the range starting at
1195  * pos for count bytes (inclusive).
1196  */
1197 static int ocfs2_check_range_for_holes(struct inode *inode, loff_t pos,
1198                                        size_t count)
1199 {
1200         int ret = 0;
1201         unsigned int extent_flags;
1202         u32 cpos, clusters, extent_len, phys_cpos;
1203         struct super_block *sb = inode->i_sb;
1204
1205         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1206         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1207
1208         while (clusters) {
1209                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1210                                          &extent_flags);
1211                 if (ret < 0) {
1212                         mlog_errno(ret);
1213                         goto out;
1214                 }
1215
1216                 if (phys_cpos == 0 || (extent_flags & OCFS2_EXT_UNWRITTEN)) {
1217                         ret = 1;
1218                         break;
1219                 }
1220
1221                 if (extent_len > clusters)
1222                         extent_len = clusters;
1223
1224                 clusters -= extent_len;
1225                 cpos += extent_len;
1226         }
1227 out:
1228         return ret;
1229 }
1230
1231 static int ocfs2_write_remove_suid(struct inode *inode)
1232 {
1233         int ret;
1234         struct buffer_head *bh = NULL;
1235
1236         ret = ocfs2_read_inode_block(inode, &bh);
1237         if (ret < 0) {
1238                 mlog_errno(ret);
1239                 goto out;
1240         }
1241
1242         ret =  __ocfs2_write_remove_suid(inode, bh);
1243 out:
1244         brelse(bh);
1245         return ret;
1246 }
1247
1248 /*
1249  * Allocate enough extents to cover the region starting at byte offset
1250  * start for len bytes. Existing extents are skipped, any extents
1251  * added are marked as "unwritten".
1252  */
1253 static int ocfs2_allocate_unwritten_extents(struct inode *inode,
1254                                             u64 start, u64 len)
1255 {
1256         int ret;
1257         u32 cpos, phys_cpos, clusters, alloc_size;
1258         u64 end = start + len;
1259         struct buffer_head *di_bh = NULL;
1260
1261         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1262                 ret = ocfs2_read_inode_block(inode, &di_bh);
1263                 if (ret) {
1264                         mlog_errno(ret);
1265                         goto out;
1266                 }
1267
1268                 /*
1269                  * Nothing to do if the requested reservation range
1270                  * fits within the inode.
1271                  */
1272                 if (ocfs2_size_fits_inline_data(di_bh, end))
1273                         goto out;
1274
1275                 ret = ocfs2_convert_inline_data_to_extents(inode, di_bh);
1276                 if (ret) {
1277                         mlog_errno(ret);
1278                         goto out;
1279                 }
1280         }
1281
1282         /*
1283          * We consider both start and len to be inclusive.
1284          */
1285         cpos = start >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1286         clusters = ocfs2_clusters_for_bytes(inode->i_sb, start + len);
1287         clusters -= cpos;
1288
1289         while (clusters) {
1290                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos,
1291                                          &alloc_size, NULL);
1292                 if (ret) {
1293                         mlog_errno(ret);
1294                         goto out;
1295                 }
1296
1297                 /*
1298                  * Hole or existing extent len can be arbitrary, so
1299                  * cap it to our own allocation request.
1300                  */
1301                 if (alloc_size > clusters)
1302                         alloc_size = clusters;
1303
1304                 if (phys_cpos) {
1305                         /*
1306                          * We already have an allocation at this
1307                          * region so we can safely skip it.
1308                          */
1309                         goto next;
1310                 }
1311
1312                 ret = __ocfs2_extend_allocation(inode, cpos, alloc_size, 1);
1313                 if (ret) {
1314                         if (ret != -ENOSPC)
1315                                 mlog_errno(ret);
1316                         goto out;
1317                 }
1318
1319 next:
1320                 cpos += alloc_size;
1321                 clusters -= alloc_size;
1322         }
1323
1324         ret = 0;
1325 out:
1326
1327         brelse(di_bh);
1328         return ret;
1329 }
1330
1331 /*
1332  * Truncate a byte range, avoiding pages within partial clusters. This
1333  * preserves those pages for the zeroing code to write to.
1334  */
1335 static void ocfs2_truncate_cluster_pages(struct inode *inode, u64 byte_start,
1336                                          u64 byte_len)
1337 {
1338         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1339         loff_t start, end;
1340         struct address_space *mapping = inode->i_mapping;
1341
1342         start = (loff_t)ocfs2_align_bytes_to_clusters(inode->i_sb, byte_start);
1343         end = byte_start + byte_len;
1344         end = end & ~(osb->s_clustersize - 1);
1345
1346         if (start < end) {
1347                 unmap_mapping_range(mapping, start, end - start, 0);
1348                 truncate_inode_pages_range(mapping, start, end - 1);
1349         }
1350 }
1351
1352 static int ocfs2_zero_partial_clusters(struct inode *inode,
1353                                        u64 start, u64 len)
1354 {
1355         int ret = 0;
1356         u64 tmpend, end = start + len;
1357         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1358         unsigned int csize = osb->s_clustersize;
1359         handle_t *handle;
1360
1361         /*
1362          * The "start" and "end" values are NOT necessarily part of
1363          * the range whose allocation is being deleted. Rather, this
1364          * is what the user passed in with the request. We must zero
1365          * partial clusters here. There's no need to worry about
1366          * physical allocation - the zeroing code knows to skip holes.
1367          */
1368         mlog(0, "byte start: %llu, end: %llu\n",
1369              (unsigned long long)start, (unsigned long long)end);
1370
1371         /*
1372          * If both edges are on a cluster boundary then there's no
1373          * zeroing required as the region is part of the allocation to
1374          * be truncated.
1375          */
1376         if ((start & (csize - 1)) == 0 && (end & (csize - 1)) == 0)
1377                 goto out;
1378
1379         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1380         if (IS_ERR(handle)) {
1381                 ret = PTR_ERR(handle);
1382                 mlog_errno(ret);
1383                 goto out;
1384         }
1385
1386         /*
1387          * We want to get the byte offset of the end of the 1st cluster.
1388          */
1389         tmpend = (u64)osb->s_clustersize + (start & ~(osb->s_clustersize - 1));
1390         if (tmpend > end)
1391                 tmpend = end;
1392
1393         mlog(0, "1st range: start: %llu, tmpend: %llu\n",
1394              (unsigned long long)start, (unsigned long long)tmpend);
1395
1396         ret = ocfs2_zero_range_for_truncate(inode, handle, start, tmpend);
1397         if (ret)
1398                 mlog_errno(ret);
1399
1400         if (tmpend < end) {
1401                 /*
1402                  * This may make start and end equal, but the zeroing
1403                  * code will skip any work in that case so there's no
1404                  * need to catch it up here.
1405                  */
1406                 start = end & ~(osb->s_clustersize - 1);
1407
1408                 mlog(0, "2nd range: start: %llu, end: %llu\n",
1409                      (unsigned long long)start, (unsigned long long)end);
1410
1411                 ret = ocfs2_zero_range_for_truncate(inode, handle, start, end);
1412                 if (ret)
1413                         mlog_errno(ret);
1414         }
1415
1416         ocfs2_commit_trans(osb, handle);
1417 out:
1418         return ret;
1419 }
1420
1421 static int ocfs2_find_rec(struct ocfs2_extent_list *el, u32 pos)
1422 {
1423         int i;
1424         struct ocfs2_extent_rec *rec = NULL;
1425
1426         for (i = le16_to_cpu(el->l_next_free_rec) - 1; i >= 0; i--) {
1427
1428                 rec = &el->l_recs[i];
1429
1430                 if (le32_to_cpu(rec->e_cpos) < pos)
1431                         break;
1432         }
1433
1434         return i;
1435 }
1436
1437 /*
1438  * Helper to calculate the punching pos and length in one run, we handle the
1439  * following three cases in order:
1440  *
1441  * - remove the entire record
1442  * - remove a partial record
1443  * - no record needs to be removed (hole-punching completed)
1444 */
1445 static void ocfs2_calc_trunc_pos(struct inode *inode,
1446                                  struct ocfs2_extent_list *el,
1447                                  struct ocfs2_extent_rec *rec,
1448                                  u32 trunc_start, u32 *trunc_cpos,
1449                                  u32 *trunc_len, u32 *trunc_end,
1450                                  u64 *blkno, int *done)
1451 {
1452         int ret = 0;
1453         u32 coff, range;
1454
1455         range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
1456
1457         if (le32_to_cpu(rec->e_cpos) >= trunc_start) {
1458                 *trunc_cpos = le32_to_cpu(rec->e_cpos);
1459                 /*
1460                  * Skip holes if any.
1461                  */
1462                 if (range < *trunc_end)
1463                         *trunc_end = range;
1464                 *trunc_len = *trunc_end - le32_to_cpu(rec->e_cpos);
1465                 *blkno = le64_to_cpu(rec->e_blkno);
1466                 *trunc_end = le32_to_cpu(rec->e_cpos);
1467         } else if (range > trunc_start) {
1468                 *trunc_cpos = trunc_start;
1469                 *trunc_len = *trunc_end - trunc_start;
1470                 coff = trunc_start - le32_to_cpu(rec->e_cpos);
1471                 *blkno = le64_to_cpu(rec->e_blkno) +
1472                                 ocfs2_clusters_to_blocks(inode->i_sb, coff);
1473                 *trunc_end = trunc_start;
1474         } else {
1475                 /*
1476                  * It may have two following possibilities:
1477                  *
1478                  * - last record has been removed
1479                  * - trunc_start was within a hole
1480                  *
1481                  * both two cases mean the completion of hole punching.
1482                  */
1483                 ret = 1;
1484         }
1485
1486         *done = ret;
1487 }
1488
1489 static int ocfs2_remove_inode_range(struct inode *inode,
1490                                     struct buffer_head *di_bh, u64 byte_start,
1491                                     u64 byte_len)
1492 {
1493         int ret = 0, flags = 0, done = 0, i;
1494         u32 trunc_start, trunc_len, trunc_end, trunc_cpos, phys_cpos;
1495         u32 cluster_in_el;
1496         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1497         struct ocfs2_cached_dealloc_ctxt dealloc;
1498         struct address_space *mapping = inode->i_mapping;
1499         struct ocfs2_extent_tree et;
1500         struct ocfs2_path *path = NULL;
1501         struct ocfs2_extent_list *el = NULL;
1502         struct ocfs2_extent_rec *rec = NULL;
1503         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1504         u64 blkno, refcount_loc = le64_to_cpu(di->i_refcount_loc);
1505
1506         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
1507         ocfs2_init_dealloc_ctxt(&dealloc);
1508
1509         if (byte_len == 0)
1510                 return 0;
1511
1512         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1513                 ret = ocfs2_truncate_inline(inode, di_bh, byte_start,
1514                                             byte_start + byte_len, 0);
1515                 if (ret) {
1516                         mlog_errno(ret);
1517                         goto out;
1518                 }
1519                 /*
1520                  * There's no need to get fancy with the page cache
1521                  * truncate of an inline-data inode. We're talking
1522                  * about less than a page here, which will be cached
1523                  * in the dinode buffer anyway.
1524                  */
1525                 unmap_mapping_range(mapping, 0, 0, 0);
1526                 truncate_inode_pages(mapping, 0);
1527                 goto out;
1528         }
1529
1530         /*
1531          * For reflinks, we may need to CoW 2 clusters which might be
1532          * partially zero'd later, if hole's start and end offset were
1533          * within one cluster(means is not exactly aligned to clustersize).
1534          */
1535
1536         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) {
1537
1538                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start);
1539                 if (ret) {
1540                         mlog_errno(ret);
1541                         goto out;
1542                 }
1543
1544                 ret = ocfs2_cow_file_pos(inode, di_bh, byte_start + byte_len);
1545                 if (ret) {
1546                         mlog_errno(ret);
1547                         goto out;
1548                 }
1549         }
1550
1551         trunc_start = ocfs2_clusters_for_bytes(osb->sb, byte_start);
1552         trunc_end = (byte_start + byte_len) >> osb->s_clustersize_bits;
1553         cluster_in_el = trunc_end;
1554
1555         mlog(0, "Inode: %llu, start: %llu, len: %llu, cstart: %u, cend: %u\n",
1556              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1557              (unsigned long long)byte_start,
1558              (unsigned long long)byte_len, trunc_start, trunc_end);
1559
1560         ret = ocfs2_zero_partial_clusters(inode, byte_start, byte_len);
1561         if (ret) {
1562                 mlog_errno(ret);
1563                 goto out;
1564         }
1565
1566         path = ocfs2_new_path_from_et(&et);
1567         if (!path) {
1568                 ret = -ENOMEM;
1569                 mlog_errno(ret);
1570                 goto out;
1571         }
1572
1573         while (trunc_end > trunc_start) {
1574
1575                 ret = ocfs2_find_path(INODE_CACHE(inode), path,
1576                                       cluster_in_el);
1577                 if (ret) {
1578                         mlog_errno(ret);
1579                         goto out;
1580                 }
1581
1582                 el = path_leaf_el(path);
1583
1584                 i = ocfs2_find_rec(el, trunc_end);
1585                 /*
1586                  * Need to go to previous extent block.
1587                  */
1588                 if (i < 0) {
1589                         if (path->p_tree_depth == 0)
1590                                 break;
1591
1592                         ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb,
1593                                                             path,
1594                                                             &cluster_in_el);
1595                         if (ret) {
1596                                 mlog_errno(ret);
1597                                 goto out;
1598                         }
1599
1600                         /*
1601                          * We've reached the leftmost extent block,
1602                          * it's safe to leave.
1603                          */
1604                         if (cluster_in_el == 0)
1605                                 break;
1606
1607                         /*
1608                          * The 'pos' searched for previous extent block is
1609                          * always one cluster less than actual trunc_end.
1610                          */
1611                         trunc_end = cluster_in_el + 1;
1612
1613                         ocfs2_reinit_path(path, 1);
1614
1615                         continue;
1616
1617                 } else
1618                         rec = &el->l_recs[i];
1619
1620                 ocfs2_calc_trunc_pos(inode, el, rec, trunc_start, &trunc_cpos,
1621                                      &trunc_len, &trunc_end, &blkno, &done);
1622                 if (done)
1623                         break;
1624
1625                 flags = rec->e_flags;
1626                 phys_cpos = ocfs2_blocks_to_clusters(inode->i_sb, blkno);
1627
1628                 ret = ocfs2_remove_btree_range(inode, &et, trunc_cpos,
1629                                                phys_cpos, trunc_len, flags,
1630                                                &dealloc, refcount_loc);
1631                 if (ret < 0) {
1632                         mlog_errno(ret);
1633                         goto out;
1634                 }
1635
1636                 cluster_in_el = trunc_end;
1637
1638                 ocfs2_reinit_path(path, 1);
1639         }
1640
1641         ocfs2_truncate_cluster_pages(inode, byte_start, byte_len);
1642
1643 out:
1644         ocfs2_schedule_truncate_log_flush(osb, 1);
1645         ocfs2_run_deallocs(osb, &dealloc);
1646
1647         return ret;
1648 }
1649
1650 /*
1651  * Parts of this function taken from xfs_change_file_space()
1652  */
1653 static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
1654                                      loff_t f_pos, unsigned int cmd,
1655                                      struct ocfs2_space_resv *sr,
1656                                      int change_size)
1657 {
1658         int ret;
1659         s64 llen;
1660         loff_t size;
1661         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1662         struct buffer_head *di_bh = NULL;
1663         handle_t *handle;
1664         unsigned long long max_off = inode->i_sb->s_maxbytes;
1665
1666         if (ocfs2_is_hard_readonly(osb) || ocfs2_is_soft_readonly(osb))
1667                 return -EROFS;
1668
1669         mutex_lock(&inode->i_mutex);
1670
1671         /*
1672          * This prevents concurrent writes on other nodes
1673          */
1674         ret = ocfs2_rw_lock(inode, 1);
1675         if (ret) {
1676                 mlog_errno(ret);
1677                 goto out;
1678         }
1679
1680         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1681         if (ret) {
1682                 mlog_errno(ret);
1683                 goto out_rw_unlock;
1684         }
1685
1686         if (inode->i_flags & (S_IMMUTABLE|S_APPEND)) {
1687                 ret = -EPERM;
1688                 goto out_inode_unlock;
1689         }
1690
1691         switch (sr->l_whence) {
1692         case 0: /*SEEK_SET*/
1693                 break;
1694         case 1: /*SEEK_CUR*/
1695                 sr->l_start += f_pos;
1696                 break;
1697         case 2: /*SEEK_END*/
1698                 sr->l_start += i_size_read(inode);
1699                 break;
1700         default:
1701                 ret = -EINVAL;
1702                 goto out_inode_unlock;
1703         }
1704         sr->l_whence = 0;
1705
1706         llen = sr->l_len > 0 ? sr->l_len - 1 : sr->l_len;
1707
1708         if (sr->l_start < 0
1709             || sr->l_start > max_off
1710             || (sr->l_start + llen) < 0
1711             || (sr->l_start + llen) > max_off) {
1712                 ret = -EINVAL;
1713                 goto out_inode_unlock;
1714         }
1715         size = sr->l_start + sr->l_len;
1716
1717         if (cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) {
1718                 if (sr->l_len <= 0) {
1719                         ret = -EINVAL;
1720                         goto out_inode_unlock;
1721                 }
1722         }
1723
1724         if (file && should_remove_suid(file->f_path.dentry)) {
1725                 ret = __ocfs2_write_remove_suid(inode, di_bh);
1726                 if (ret) {
1727                         mlog_errno(ret);
1728                         goto out_inode_unlock;
1729                 }
1730         }
1731
1732         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1733         switch (cmd) {
1734         case OCFS2_IOC_RESVSP:
1735         case OCFS2_IOC_RESVSP64:
1736                 /*
1737                  * This takes unsigned offsets, but the signed ones we
1738                  * pass have been checked against overflow above.
1739                  */
1740                 ret = ocfs2_allocate_unwritten_extents(inode, sr->l_start,
1741                                                        sr->l_len);
1742                 break;
1743         case OCFS2_IOC_UNRESVSP:
1744         case OCFS2_IOC_UNRESVSP64:
1745                 ret = ocfs2_remove_inode_range(inode, di_bh, sr->l_start,
1746                                                sr->l_len);
1747                 break;
1748         default:
1749                 ret = -EINVAL;
1750         }
1751         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1752         if (ret) {
1753                 mlog_errno(ret);
1754                 goto out_inode_unlock;
1755         }
1756
1757         /*
1758          * We update c/mtime for these changes
1759          */
1760         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1761         if (IS_ERR(handle)) {
1762                 ret = PTR_ERR(handle);
1763                 mlog_errno(ret);
1764                 goto out_inode_unlock;
1765         }
1766
1767         if (change_size && i_size_read(inode) < size)
1768                 i_size_write(inode, size);
1769
1770         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1771         ret = ocfs2_mark_inode_dirty(handle, inode, di_bh);
1772         if (ret < 0)
1773                 mlog_errno(ret);
1774
1775         ocfs2_commit_trans(osb, handle);
1776
1777 out_inode_unlock:
1778         brelse(di_bh);
1779         ocfs2_inode_unlock(inode, 1);
1780 out_rw_unlock:
1781         ocfs2_rw_unlock(inode, 1);
1782
1783 out:
1784         mutex_unlock(&inode->i_mutex);
1785         return ret;
1786 }
1787
1788 int ocfs2_change_file_space(struct file *file, unsigned int cmd,
1789                             struct ocfs2_space_resv *sr)
1790 {
1791         struct inode *inode = file->f_path.dentry->d_inode;
1792         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1793
1794         if ((cmd == OCFS2_IOC_RESVSP || cmd == OCFS2_IOC_RESVSP64) &&
1795             !ocfs2_writes_unwritten_extents(osb))
1796                 return -ENOTTY;
1797         else if ((cmd == OCFS2_IOC_UNRESVSP || cmd == OCFS2_IOC_UNRESVSP64) &&
1798                  !ocfs2_sparse_alloc(osb))
1799                 return -ENOTTY;
1800
1801         if (!S_ISREG(inode->i_mode))
1802                 return -EINVAL;
1803
1804         if (!(file->f_mode & FMODE_WRITE))
1805                 return -EBADF;
1806
1807         return __ocfs2_change_file_space(file, inode, file->f_pos, cmd, sr, 0);
1808 }
1809
1810 static long ocfs2_fallocate(struct inode *inode, int mode, loff_t offset,
1811                             loff_t len)
1812 {
1813         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1814         struct ocfs2_space_resv sr;
1815         int change_size = 1;
1816
1817         if (!ocfs2_writes_unwritten_extents(osb))
1818                 return -EOPNOTSUPP;
1819
1820         if (S_ISDIR(inode->i_mode))
1821                 return -ENODEV;
1822
1823         if (mode & FALLOC_FL_KEEP_SIZE)
1824                 change_size = 0;
1825
1826         sr.l_whence = 0;
1827         sr.l_start = (s64)offset;
1828         sr.l_len = (s64)len;
1829
1830         return __ocfs2_change_file_space(NULL, inode, offset,
1831                                          OCFS2_IOC_RESVSP64, &sr, change_size);
1832 }
1833
1834 int ocfs2_check_range_for_refcount(struct inode *inode, loff_t pos,
1835                                    size_t count)
1836 {
1837         int ret = 0;
1838         unsigned int extent_flags;
1839         u32 cpos, clusters, extent_len, phys_cpos;
1840         struct super_block *sb = inode->i_sb;
1841
1842         if (!ocfs2_refcount_tree(OCFS2_SB(inode->i_sb)) ||
1843             !(OCFS2_I(inode)->ip_dyn_features & OCFS2_HAS_REFCOUNT_FL) ||
1844             OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
1845                 return 0;
1846
1847         cpos = pos >> OCFS2_SB(sb)->s_clustersize_bits;
1848         clusters = ocfs2_clusters_for_bytes(sb, pos + count) - cpos;
1849
1850         while (clusters) {
1851                 ret = ocfs2_get_clusters(inode, cpos, &phys_cpos, &extent_len,
1852                                          &extent_flags);
1853                 if (ret < 0) {
1854                         mlog_errno(ret);
1855                         goto out;
1856                 }
1857
1858                 if (phys_cpos && (extent_flags & OCFS2_EXT_REFCOUNTED)) {
1859                         ret = 1;
1860                         break;
1861                 }
1862
1863                 if (extent_len > clusters)
1864                         extent_len = clusters;
1865
1866                 clusters -= extent_len;
1867                 cpos += extent_len;
1868         }
1869 out:
1870         return ret;
1871 }
1872
1873 static int ocfs2_prepare_inode_for_refcount(struct inode *inode,
1874                                             loff_t pos, size_t count,
1875                                             int *meta_level)
1876 {
1877         int ret;
1878         struct buffer_head *di_bh = NULL;
1879         u32 cpos = pos >> OCFS2_SB(inode->i_sb)->s_clustersize_bits;
1880         u32 clusters =
1881                 ocfs2_clusters_for_bytes(inode->i_sb, pos + count) - cpos;
1882
1883         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1884         if (ret) {
1885                 mlog_errno(ret);
1886                 goto out;
1887         }
1888
1889         *meta_level = 1;
1890
1891         ret = ocfs2_refcount_cow(inode, di_bh, cpos, clusters, UINT_MAX);
1892         if (ret)
1893                 mlog_errno(ret);
1894 out:
1895         brelse(di_bh);
1896         return ret;
1897 }
1898
1899 static int ocfs2_prepare_inode_for_write(struct dentry *dentry,
1900                                          loff_t *ppos,
1901                                          size_t count,
1902                                          int appending,
1903                                          int *direct_io,
1904                                          int *has_refcount)
1905 {
1906         int ret = 0, meta_level = 0;
1907         struct inode *inode = dentry->d_inode;
1908         loff_t saved_pos, end;
1909
1910         /*
1911          * We start with a read level meta lock and only jump to an ex
1912          * if we need to make modifications here.
1913          */
1914         for(;;) {
1915                 ret = ocfs2_inode_lock(inode, NULL, meta_level);
1916                 if (ret < 0) {
1917                         meta_level = -1;
1918                         mlog_errno(ret);
1919                         goto out;
1920                 }
1921
1922                 /* Clear suid / sgid if necessary. We do this here
1923                  * instead of later in the write path because
1924                  * remove_suid() calls ->setattr without any hint that
1925                  * we may have already done our cluster locking. Since
1926                  * ocfs2_setattr() *must* take cluster locks to
1927                  * proceeed, this will lead us to recursively lock the
1928                  * inode. There's also the dinode i_size state which
1929                  * can be lost via setattr during extending writes (we
1930                  * set inode->i_size at the end of a write. */
1931                 if (should_remove_suid(dentry)) {
1932                         if (meta_level == 0) {
1933                                 ocfs2_inode_unlock(inode, meta_level);
1934                                 meta_level = 1;
1935                                 continue;
1936                         }
1937
1938                         ret = ocfs2_write_remove_suid(inode);
1939                         if (ret < 0) {
1940                                 mlog_errno(ret);
1941                                 goto out_unlock;
1942                         }
1943                 }
1944
1945                 /* work on a copy of ppos until we're sure that we won't have
1946                  * to recalculate it due to relocking. */
1947                 if (appending) {
1948                         saved_pos = i_size_read(inode);
1949                         mlog(0, "O_APPEND: inode->i_size=%llu\n", saved_pos);
1950                 } else {
1951                         saved_pos = *ppos;
1952                 }
1953
1954                 end = saved_pos + count;
1955
1956                 ret = ocfs2_check_range_for_refcount(inode, saved_pos, count);
1957                 if (ret == 1) {
1958                         ocfs2_inode_unlock(inode, meta_level);
1959                         meta_level = -1;
1960
1961                         ret = ocfs2_prepare_inode_for_refcount(inode,
1962                                                                saved_pos,
1963                                                                count,
1964                                                                &meta_level);
1965                         if (has_refcount)
1966                                 *has_refcount = 1;
1967                         if (direct_io)
1968                                 *direct_io = 0;
1969                 }
1970
1971                 if (ret < 0) {
1972                         mlog_errno(ret);
1973                         goto out_unlock;
1974                 }
1975
1976                 /*
1977                  * Skip the O_DIRECT checks if we don't need
1978                  * them.
1979                  */
1980                 if (!direct_io || !(*direct_io))
1981                         break;
1982
1983                 /*
1984                  * There's no sane way to do direct writes to an inode
1985                  * with inline data.
1986                  */
1987                 if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1988                         *direct_io = 0;
1989                         break;
1990                 }
1991
1992                 /*
1993                  * Allowing concurrent direct writes means
1994                  * i_size changes wouldn't be synchronized, so
1995                  * one node could wind up truncating another
1996                  * nodes writes.
1997                  */
1998                 if (end > i_size_read(inode)) {
1999                         *direct_io = 0;
2000                         break;
2001                 }
2002
2003                 /*
2004                  * We don't fill holes during direct io, so
2005                  * check for them here. If any are found, the
2006                  * caller will have to retake some cluster
2007                  * locks and initiate the io as buffered.
2008                  */
2009                 ret = ocfs2_check_range_for_holes(inode, saved_pos, count);
2010                 if (ret == 1) {
2011                         *direct_io = 0;
2012                         ret = 0;
2013                 } else if (ret < 0)
2014                         mlog_errno(ret);
2015                 break;
2016         }
2017
2018         if (appending)
2019                 *ppos = saved_pos;
2020
2021 out_unlock:
2022         if (meta_level >= 0)
2023                 ocfs2_inode_unlock(inode, meta_level);
2024
2025 out:
2026         return ret;
2027 }
2028
2029 static ssize_t ocfs2_file_aio_write(struct kiocb *iocb,
2030                                     const struct iovec *iov,
2031                                     unsigned long nr_segs,
2032                                     loff_t pos)
2033 {
2034         int ret, direct_io, appending, rw_level, have_alloc_sem  = 0;
2035         int can_do_direct, has_refcount = 0;
2036         ssize_t written = 0;
2037         size_t ocount;          /* original count */
2038         size_t count;           /* after file limit checks */
2039         loff_t old_size, *ppos = &iocb->ki_pos;
2040         u32 old_clusters;
2041         struct file *file = iocb->ki_filp;
2042         struct inode *inode = file->f_path.dentry->d_inode;
2043         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2044
2045         mlog_entry("(0x%p, %u, '%.*s')\n", file,
2046                    (unsigned int)nr_segs,
2047                    file->f_path.dentry->d_name.len,
2048                    file->f_path.dentry->d_name.name);
2049
2050         if (iocb->ki_left == 0)
2051                 return 0;
2052
2053         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2054
2055         appending = file->f_flags & O_APPEND ? 1 : 0;
2056         direct_io = file->f_flags & O_DIRECT ? 1 : 0;
2057
2058         mutex_lock(&inode->i_mutex);
2059
2060 relock:
2061         /* to match setattr's i_mutex -> i_alloc_sem -> rw_lock ordering */
2062         if (direct_io) {
2063                 down_read(&inode->i_alloc_sem);
2064                 have_alloc_sem = 1;
2065         }
2066
2067         /* concurrent O_DIRECT writes are allowed */
2068         rw_level = !direct_io;
2069         ret = ocfs2_rw_lock(inode, rw_level);
2070         if (ret < 0) {
2071                 mlog_errno(ret);
2072                 goto out_sems;
2073         }
2074
2075         can_do_direct = direct_io;
2076         ret = ocfs2_prepare_inode_for_write(file->f_path.dentry, ppos,
2077                                             iocb->ki_left, appending,
2078                                             &can_do_direct, &has_refcount);
2079         if (ret < 0) {
2080                 mlog_errno(ret);
2081                 goto out;
2082         }
2083
2084         /*
2085          * We can't complete the direct I/O as requested, fall back to
2086          * buffered I/O.
2087          */
2088         if (direct_io && !can_do_direct) {
2089                 ocfs2_rw_unlock(inode, rw_level);
2090                 up_read(&inode->i_alloc_sem);
2091
2092                 have_alloc_sem = 0;
2093                 rw_level = -1;
2094
2095                 direct_io = 0;
2096                 goto relock;
2097         }
2098
2099         /*
2100          * To later detect whether a journal commit for sync writes is
2101          * necessary, we sample i_size, and cluster count here.
2102          */
2103         old_size = i_size_read(inode);
2104         old_clusters = OCFS2_I(inode)->ip_clusters;
2105
2106         /* communicate with ocfs2_dio_end_io */
2107         ocfs2_iocb_set_rw_locked(iocb, rw_level);
2108
2109         if (direct_io) {
2110                 ret = generic_segment_checks(iov, &nr_segs, &ocount,
2111                                              VERIFY_READ);
2112                 if (ret)
2113                         goto out_dio;
2114
2115                 count = ocount;
2116                 ret = generic_write_checks(file, ppos, &count,
2117                                            S_ISBLK(inode->i_mode));
2118                 if (ret)
2119                         goto out_dio;
2120
2121                 written = generic_file_direct_write(iocb, iov, &nr_segs, *ppos,
2122                                                     ppos, count, ocount);
2123                 if (written < 0) {
2124                         /*
2125                          * direct write may have instantiated a few
2126                          * blocks outside i_size. Trim these off again.
2127                          * Don't need i_size_read because we hold i_mutex.
2128                          */
2129                         if (*ppos + count > inode->i_size)
2130                                 vmtruncate(inode, inode->i_size);
2131                         ret = written;
2132                         goto out_dio;
2133                 }
2134         } else {
2135                 written = __generic_file_aio_write(iocb, iov, nr_segs, ppos);
2136         }
2137
2138 out_dio:
2139         /* buffered aio wouldn't have proper lock coverage today */
2140         BUG_ON(ret == -EIOCBQUEUED && !(file->f_flags & O_DIRECT));
2141
2142         if (((file->f_flags & O_DSYNC) && !direct_io) || IS_SYNC(inode) ||
2143             ((file->f_flags & O_DIRECT) && has_refcount)) {
2144                 ret = filemap_fdatawrite_range(file->f_mapping, pos,
2145                                                pos + count - 1);
2146                 if (ret < 0)
2147                         written = ret;
2148
2149                 if (!ret && (old_size != i_size_read(inode) ||
2150                     old_clusters != OCFS2_I(inode)->ip_clusters ||
2151                     has_refcount)) {
2152                         ret = jbd2_journal_force_commit(osb->journal->j_journal);
2153                         if (ret < 0)
2154                                 written = ret;
2155                 }
2156
2157                 if (!ret)
2158                         ret = filemap_fdatawait_range(file->f_mapping, pos,
2159                                                       pos + count - 1);
2160         }
2161
2162         /*
2163          * deep in g_f_a_w_n()->ocfs2_direct_IO we pass in a ocfs2_dio_end_io
2164          * function pointer which is called when o_direct io completes so that
2165          * it can unlock our rw lock.  (it's the clustered equivalent of
2166          * i_alloc_sem; protects truncate from racing with pending ios).
2167          * Unfortunately there are error cases which call end_io and others
2168          * that don't.  so we don't have to unlock the rw_lock if either an
2169          * async dio is going to do it in the future or an end_io after an
2170          * error has already done it.
2171          */
2172         if ((ret == -EIOCBQUEUED) || (!ocfs2_iocb_is_rw_locked(iocb))) {
2173                 rw_level = -1;
2174                 have_alloc_sem = 0;
2175         }
2176
2177 out:
2178         if (rw_level != -1)
2179                 ocfs2_rw_unlock(inode, rw_level);
2180
2181 out_sems:
2182         if (have_alloc_sem)
2183                 up_read(&inode->i_alloc_sem);
2184
2185         mutex_unlock(&inode->i_mutex);
2186
2187         if (written)
2188                 ret = written;
2189         mlog_exit(ret);
2190         return ret;
2191 }
2192
2193 static int ocfs2_splice_to_file(struct pipe_inode_info *pipe,
2194                                 struct file *out,
2195                                 struct splice_desc *sd)
2196 {
2197         int ret;
2198
2199         ret = ocfs2_prepare_inode_for_write(out->f_path.dentry, &sd->pos,
2200                                             sd->total_len, 0, NULL, NULL);
2201         if (ret < 0) {
2202                 mlog_errno(ret);
2203                 return ret;
2204         }
2205
2206         return splice_from_pipe_feed(pipe, sd, pipe_to_file);
2207 }
2208
2209 static ssize_t ocfs2_file_splice_write(struct pipe_inode_info *pipe,
2210                                        struct file *out,
2211                                        loff_t *ppos,
2212                                        size_t len,
2213                                        unsigned int flags)
2214 {
2215         int ret;
2216         struct address_space *mapping = out->f_mapping;
2217         struct inode *inode = mapping->host;
2218         struct splice_desc sd = {
2219                 .total_len = len,
2220                 .flags = flags,
2221                 .pos = *ppos,
2222                 .u.file = out,
2223         };
2224
2225         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", out, pipe,
2226                    (unsigned int)len,
2227                    out->f_path.dentry->d_name.len,
2228                    out->f_path.dentry->d_name.name);
2229
2230         if (pipe->inode)
2231                 mutex_lock_nested(&pipe->inode->i_mutex, I_MUTEX_PARENT);
2232
2233         splice_from_pipe_begin(&sd);
2234         do {
2235                 ret = splice_from_pipe_next(pipe, &sd);
2236                 if (ret <= 0)
2237                         break;
2238
2239                 mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
2240                 ret = ocfs2_rw_lock(inode, 1);
2241                 if (ret < 0)
2242                         mlog_errno(ret);
2243                 else {
2244                         ret = ocfs2_splice_to_file(pipe, out, &sd);
2245                         ocfs2_rw_unlock(inode, 1);
2246                 }
2247                 mutex_unlock(&inode->i_mutex);
2248         } while (ret > 0);
2249         splice_from_pipe_end(pipe, &sd);
2250
2251         if (pipe->inode)
2252                 mutex_unlock(&pipe->inode->i_mutex);
2253
2254         if (sd.num_spliced)
2255                 ret = sd.num_spliced;
2256
2257         if (ret > 0) {
2258                 unsigned long nr_pages;
2259                 int err;
2260
2261                 nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
2262
2263                 err = generic_write_sync(out, *ppos, ret);
2264                 if (err)
2265                         ret = err;
2266                 else
2267                         *ppos += ret;
2268
2269                 balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
2270         }
2271
2272         mlog_exit(ret);
2273         return ret;
2274 }
2275
2276 static ssize_t ocfs2_file_splice_read(struct file *in,
2277                                       loff_t *ppos,
2278                                       struct pipe_inode_info *pipe,
2279                                       size_t len,
2280                                       unsigned int flags)
2281 {
2282         int ret = 0, lock_level = 0;
2283         struct inode *inode = in->f_path.dentry->d_inode;
2284
2285         mlog_entry("(0x%p, 0x%p, %u, '%.*s')\n", in, pipe,
2286                    (unsigned int)len,
2287                    in->f_path.dentry->d_name.len,
2288                    in->f_path.dentry->d_name.name);
2289
2290         /*
2291          * See the comment in ocfs2_file_aio_read()
2292          */
2293         ret = ocfs2_inode_lock_atime(inode, in->f_vfsmnt, &lock_level);
2294         if (ret < 0) {
2295                 mlog_errno(ret);
2296                 goto bail;
2297         }
2298         ocfs2_inode_unlock(inode, lock_level);
2299
2300         ret = generic_file_splice_read(in, ppos, pipe, len, flags);
2301
2302 bail:
2303         mlog_exit(ret);
2304         return ret;
2305 }
2306
2307 static ssize_t ocfs2_file_aio_read(struct kiocb *iocb,
2308                                    const struct iovec *iov,
2309                                    unsigned long nr_segs,
2310                                    loff_t pos)
2311 {
2312         int ret = 0, rw_level = -1, have_alloc_sem = 0, lock_level = 0;
2313         struct file *filp = iocb->ki_filp;
2314         struct inode *inode = filp->f_path.dentry->d_inode;
2315
2316         mlog_entry("(0x%p, %u, '%.*s')\n", filp,
2317                    (unsigned int)nr_segs,
2318                    filp->f_path.dentry->d_name.len,
2319                    filp->f_path.dentry->d_name.name);
2320
2321         if (!inode) {
2322                 ret = -EINVAL;
2323                 mlog_errno(ret);
2324                 goto bail;
2325         }
2326
2327         /*
2328          * buffered reads protect themselves in ->readpage().  O_DIRECT reads
2329          * need locks to protect pending reads from racing with truncate.
2330          */
2331         if (filp->f_flags & O_DIRECT) {
2332                 down_read(&inode->i_alloc_sem);
2333                 have_alloc_sem = 1;
2334
2335                 ret = ocfs2_rw_lock(inode, 0);
2336                 if (ret < 0) {
2337                         mlog_errno(ret);
2338                         goto bail;
2339                 }
2340                 rw_level = 0;
2341                 /* communicate with ocfs2_dio_end_io */
2342                 ocfs2_iocb_set_rw_locked(iocb, rw_level);
2343         }
2344
2345         /*
2346          * We're fine letting folks race truncates and extending
2347          * writes with read across the cluster, just like they can
2348          * locally. Hence no rw_lock during read.
2349          *
2350          * Take and drop the meta data lock to update inode fields
2351          * like i_size. This allows the checks down below
2352          * generic_file_aio_read() a chance of actually working.
2353          */
2354         ret = ocfs2_inode_lock_atime(inode, filp->f_vfsmnt, &lock_level);
2355         if (ret < 0) {
2356                 mlog_errno(ret);
2357                 goto bail;
2358         }
2359         ocfs2_inode_unlock(inode, lock_level);
2360
2361         ret = generic_file_aio_read(iocb, iov, nr_segs, iocb->ki_pos);
2362         if (ret == -EINVAL)
2363                 mlog(0, "generic_file_aio_read returned -EINVAL\n");
2364
2365         /* buffered aio wouldn't have proper lock coverage today */
2366         BUG_ON(ret == -EIOCBQUEUED && !(filp->f_flags & O_DIRECT));
2367
2368         /* see ocfs2_file_aio_write */
2369         if (ret == -EIOCBQUEUED || !ocfs2_iocb_is_rw_locked(iocb)) {
2370                 rw_level = -1;
2371                 have_alloc_sem = 0;
2372         }
2373
2374 bail:
2375         if (have_alloc_sem)
2376                 up_read(&inode->i_alloc_sem);
2377         if (rw_level != -1)
2378                 ocfs2_rw_unlock(inode, rw_level);
2379         mlog_exit(ret);
2380
2381         return ret;
2382 }
2383
2384 const struct inode_operations ocfs2_file_iops = {
2385         .setattr        = ocfs2_setattr,
2386         .getattr        = ocfs2_getattr,
2387         .permission     = ocfs2_permission,
2388         .setxattr       = generic_setxattr,
2389         .getxattr       = generic_getxattr,
2390         .listxattr      = ocfs2_listxattr,
2391         .removexattr    = generic_removexattr,
2392         .fallocate      = ocfs2_fallocate,
2393         .fiemap         = ocfs2_fiemap,
2394 };
2395
2396 const struct inode_operations ocfs2_special_file_iops = {
2397         .setattr        = ocfs2_setattr,
2398         .getattr        = ocfs2_getattr,
2399         .permission     = ocfs2_permission,
2400 };
2401
2402 /*
2403  * Other than ->lock, keep ocfs2_fops and ocfs2_dops in sync with
2404  * ocfs2_fops_no_plocks and ocfs2_dops_no_plocks!
2405  */
2406 const struct file_operations ocfs2_fops = {
2407         .llseek         = generic_file_llseek,
2408         .read           = do_sync_read,
2409         .write          = do_sync_write,
2410         .mmap           = ocfs2_mmap,
2411         .fsync          = ocfs2_sync_file,
2412         .release        = ocfs2_file_release,
2413         .open           = ocfs2_file_open,
2414         .aio_read       = ocfs2_file_aio_read,
2415         .aio_write      = ocfs2_file_aio_write,
2416         .unlocked_ioctl = ocfs2_ioctl,
2417 #ifdef CONFIG_COMPAT
2418         .compat_ioctl   = ocfs2_compat_ioctl,
2419 #endif
2420         .lock           = ocfs2_lock,
2421         .flock          = ocfs2_flock,
2422         .splice_read    = ocfs2_file_splice_read,
2423         .splice_write   = ocfs2_file_splice_write,
2424 };
2425
2426 const struct file_operations ocfs2_dops = {
2427         .llseek         = generic_file_llseek,
2428         .read           = generic_read_dir,
2429         .readdir        = ocfs2_readdir,
2430         .fsync          = ocfs2_sync_file,
2431         .release        = ocfs2_dir_release,
2432         .open           = ocfs2_dir_open,
2433         .unlocked_ioctl = ocfs2_ioctl,
2434 #ifdef CONFIG_COMPAT
2435         .compat_ioctl   = ocfs2_compat_ioctl,
2436 #endif
2437         .lock           = ocfs2_lock,
2438         .flock          = ocfs2_flock,
2439 };
2440
2441 /*
2442  * POSIX-lockless variants of our file_operations.
2443  *
2444  * These will be used if the underlying cluster stack does not support
2445  * posix file locking, if the user passes the "localflocks" mount
2446  * option, or if we have a local-only fs.
2447  *
2448  * ocfs2_flock is in here because all stacks handle UNIX file locks,
2449  * so we still want it in the case of no stack support for
2450  * plocks. Internally, it will do the right thing when asked to ignore
2451  * the cluster.
2452  */
2453 const struct file_operations ocfs2_fops_no_plocks = {
2454         .llseek         = generic_file_llseek,
2455         .read           = do_sync_read,
2456         .write          = do_sync_write,
2457         .mmap           = ocfs2_mmap,
2458         .fsync          = ocfs2_sync_file,
2459         .release        = ocfs2_file_release,
2460         .open           = ocfs2_file_open,
2461         .aio_read       = ocfs2_file_aio_read,
2462         .aio_write      = ocfs2_file_aio_write,
2463         .unlocked_ioctl = ocfs2_ioctl,
2464 #ifdef CONFIG_COMPAT
2465         .compat_ioctl   = ocfs2_compat_ioctl,
2466 #endif
2467         .flock          = ocfs2_flock,
2468         .splice_read    = ocfs2_file_splice_read,
2469         .splice_write   = ocfs2_file_splice_write,
2470 };
2471
2472 const struct file_operations ocfs2_dops_no_plocks = {
2473         .llseek         = generic_file_llseek,
2474         .read           = generic_read_dir,
2475         .readdir        = ocfs2_readdir,
2476         .fsync          = ocfs2_sync_file,
2477         .release        = ocfs2_dir_release,
2478         .open           = ocfs2_dir_open,
2479         .unlocked_ioctl = ocfs2_ioctl,
2480 #ifdef CONFIG_COMPAT
2481         .compat_ioctl   = ocfs2_compat_ioctl,
2482 #endif
2483         .flock          = ocfs2_flock,
2484 };